Do not check for undefined behavior in abs_diff #842
Conversation
Similar to sycl::abs, sycl::abs_diff now has undefined behavior for unrepresentable results. This patch changes the generated tests for abs_diff to avoid UB, similar to how it was done for abs in KhronosGroup#831. Signed-off-by: Larsen, Steffen <[email protected]>
I worry we would run into the same problem causing this problem when computing |
|
I mean for auto au = static_cast<U>(a) + 0x80;
auto bu = static_cast<U>(b) + 0x80;
auto h = (au > bu) ? au : bu;
auto l = (au <= bu) ? au : bu;
return h - l;So technically |
|
Is something like https://godbolt.org/z/ejbK9qPdj what you are thinking? |
| T h = (a > b) ? a : b; | ||
| T l = (a <= b) ? a : b; | ||
| return h - l; | ||
| // Using two's-complement and that unsigned integer underflow is defined as | ||
| // modulo 2^n we get the result by computing the distance based on signed |
There was a problem hiding this comment.
Thanks, now I understand the code. :-)
But is it signed? It looks like it uses unsigned comparison.
There was a problem hiding this comment.
I'm not sure why Steffen says "signed", but would it be enough to just drop this word from the comment to merge it?
P.S. Steffen is on vacation, but we need this PR.
Similar to sycl::abs, sycl::abs_diff now has undefined behavior for unrepresentable results. This patch changes the generated tests for abs_diff to avoid UB, similar to how it was done for abs in #831.